If you’ve read anything about the Model Context Protocol in the last two weeks, you may have come across a specific claim: that on August 18, Anthropic and OpenAI jointly published an “MCP v2.0” specification draft on GitHub, introducing bidirectional tool calling — MCP servers requesting capabilities from clients instead of only responding to them — with a formal RFC targeted for Q4 2026 and the community server registry crossing 2,000 entries.
It’s a specific, plausible-sounding, well-formed claim. It’s also not true, as far as any primary source shows.
What’s actually in the primary sources#
MCP’s spec repository (github.com/modelcontextprotocol/modelcontextprotocol) doesn’t version by number at all — it versions by date. The current schema lives at /schema/2026-07-28/schema.ts, following the same 2025-06-18-style dating convention the project has used since its earliest releases. There is no “2.0” tag, release, or open pull request matching that description in the repository. There’s no OpenAI co-authorship visible anywhere in the repo’s history; the listed creators are Anthropic’s David Soria Parra and Justin Spahr-Summers, the same maintainers behind the project since launch.
The protocol’s own blog (blog.modelcontextprotocol.io) tells the same story. Its most recent posts as of this writing are “The New MCP Roadmap” (August 22, authored by Soria Parra and Den Delimarsky), “The 2026-07-28 Specification” (July 28), and a note on the Ruby SDK reaching 1.0 (July 27). This blog covered that roadmap post the day after it published — five priority areas for the next spec cycle, led by agent identity and enterprise authorization, with no dates attached to any of them. Nowhere in that roadmap, which is precisely the document you’d expect to mention a major version bump if one were coming, is there any reference to a “v2.0,” a joint OpenAI initiative, bidirectional tool calling, or a Q4 RFC target.
Where the claim actually came from#
The specific “August 18, jointly published, bidirectional tool calling” framing traces to a single identifiable source: an August 24 newsletter post on bighatgroup.com titled “OpenClaw Weekly: Cloudways GA and the MCP 2.0 Draft.” The post asserts the joint-publication claim in a single sentence, references “modelcontextprotocol/specification” by name, and — this is the part worth sitting with — includes no hyperlink to the repository, the alleged PR, or any announcement from either company. It’s an assertion with a citation-shaped mention and no citation.
From there, the claim propagated the way these things do: at least two other blogs (Context Studios’ “MCP v2 Beta” post and an explainx.ai piece on an unrelated “Model Hardware Standard” research preview) picked up variations of the same framing, each adding its own gloss — a “2,000 servers, up from 400” growth stat here, a “wait for the final spec” caveat there — without any of them tracing back to a primary Anthropic or OpenAI statement. Each restatement made the claim look more corroborated, because it now appeared in more places, even though every one of those places ultimately points back to the same uncited newsletter post.
This is a specific and recognizable failure mode in AI-tooling coverage: a plausible technical claim, stated confidently enough that it reads like reporting rather than speculation, gets laundered into apparent consensus purely through repetition across secondary blogs that don’t independently verify it against the primary source. The claim isn’t wild — bidirectional tool calling and OpenAI/Anthropic protocol convergence are both things people in this space could reasonably expect to happen eventually. That plausibility is exactly what makes it spread further before anyone checks.
Why this matters beyond one rumor#
MCP’s real trajectory this year has been unglamorous but well-documented: the project ships dated spec revisions, not marketing-friendly major versions, and its actual August news — the roadmap post naming agent identity and OAuth-for-unattended-agents as the top priority for the next cycle — is a more interesting and more real story than the fabricated one. If you’re building on MCP and making architecture decisions based on an anticipated “v2.0 bidirectional tool calling” capability landing this quarter, you’re planning against a feature that doesn’t exist in any primary-source roadmap, spec draft, or announcement from either company named in the claim.
The practical takeaway for anyone tracking protocol or model news in this space: a claim naming two specific companies, a specific date, and a specific technical capability is not automatically more credible than a vague one — it’s just more specific about what to go check. When a claim like this crosses your feed, the primary spec repo and the project’s own blog are a five-minute fetch away. In this case, that five minutes was the difference between reporting real news and repeating a newsletter’s uncited guess.
Sources:
